Activity Log - Demo Research
| Activity Weekly |
4/26/2022 - 4/29/2022
|
| Datasets | Total (items) |
| Low exposure | 134 |
| Normal exposure | 381 |
| Over exposure | 224 |
| Total | 739 |
| Results b : |
| Model_1 |
Xception
|
||
| Model_2 |
VGG-16 |
||
| Model_3 |
VGG-19 |
||
| Model_4 |
ResNet50 |
||
| Model_5 |
ResNet101 |
||
| Model_6 |
MobileNetv2 |
| Solving_1 |
Regularization
|
||
| Solving_2 |
Weight Initialization
|
||
| Solving_3 |
Dropout Regularization
|
||
| Solving_4 |
Weight Constraints
|
==============================================================
For Read - Bright & Sharp - Make Sample Dataset for Training
Why? :
- Terdapat datatest foto yang kelebihan&kekurangan cahaya.
- Terdapat datatest foto yang blur.
Plan? :
- Membuat Model sederhana untuk Bright Adjusting
- Membuat Model sederhana untuk Sharp Adjusting
Implement? :
- Bright Adjusting
- pengambilan sample masih bermain di pixel, belum didata real
- soon
- Sharp Adjusting
- pengambilan sample masih bermain di pixel, belum didata real
- soon
Milestone? :
- Bright Adjusting
- Mendapatkan output bright adjust yang sesuai pada foto datatest
- Sharp Adjusting
- Mendapatkan output Sharp adjust yang sesuai pada foto datatest
Referensi Paper :
- Sharpness and Brightness Quality Assessment of Face Images for Recognition.pdf
- Contrast Enhancement of Brightness-Distorted Images by Improved Adaptive Gamma Correction.pdf
- Face detection and the effect of contrast and brightness.pdf
- Haze Image Recognition Based on Brightness Optimization Feedback and Color Correction.pdf
- Contrast and brightness balance in image enhancement using Cuckoo Search-optimized image fusion.pdf
==============================================================
For Read - Bright- Using method Normalize, ImageDataGenerator, Adaptive Gamma Correction, Histogram Clipping
Normalize :
ImageDataGenerator 1 :
ImageDataGenerator 2 :
Adaptive Gamma Correciton::
XRM4WGKU - Vidia Adistia - 3674064212900001: Bright Image
XRM7DPIC - Mike noviani - 3174046904941003: Dimmed
XRNP3RAA - Ari Yunita - 3171044406840002: Bright Image
XRO4KQ7J - Lissa Nirmala - 3273194909890001: Dimmed
XROKEC4T - Tasha Kamarita - 3171085112971001: Bright Image
XRQ3CRGM - Hilda Presti Deviyanti - 3172044507820009: Bright Image
XRQGYRGB - Liesdha Nurfitrina - 3175054210880007: Dimmed
XRQLNIT6 - Yani Ratnasari - 3173074612900001: Dimmed
Input :
Output :
Histogram Clipping :
1. automatic detect :
alpha 1.186046511627907
beta -45.06976744186047
2. automatic detect :
alpha 1.9921875
beta -209.1796875
==============================================================
For Read - Discontinued log - Brightness with Object Detection
Training and Test Data for detection
| Model | Compare with 2 model input image : |
| ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8 | SSD300: 300×300 input image, lower resolution, faster. |
| SSD512: 512×512 input image, higher resolution, more accurate. |
| Dataset image annotations | Class : |
| Lowlight, Normal, Overexpourse |
| Training Method | Setup Paths |
| install tensorflow model zoo (ssd_mobilenet_v2_fpnlite_320x320_coco17_tpu-8) | |
| install tensorflow Object Detection API | |
| Create label map (Checkpoint) | |
| Create Tensorflow Records (labelmap.pbtxt , test.records, train.records) | |
| Copy Model Config to Training Folder | |
| Update Config for Transfer Learning | |
| Setup pipeline.config | |
| Train the model | |
| Evaluate the model | |
| Load Train Model from Checkpoint | |
| Detect from image |
| Pipeline Config Paramater | Model SSD MobileNet V2: | Value |
| num_classes | 3 | |
| fixed_shape_resizer | h:320 w:320 | |
| feature extractor | ssd_mobilenet_v2_fpn_keras | |
| box_coder | faster_rcnn_box_coder | |
| steps | 10 | |
| batch_size | 4 |
Evaluation training has not reached the desired quality
| Tensorboard train evaluation |
| Result of Detection | |
Confusing matrix with tensorboard --logdir=.
| Confusion Matrix | |
| AP/AR | IOU | Area | MaxDets | Value |
| Average Precision | 0.50:0.95 | all | 100 | 0.593 |
| 0.5 | all | 100 | 0.901 | |
| 0.75 | all | 100 | 0.513 | |
| 0.50:0.95 | small | 100 | -1 | |
| 0.50:0.95 | medium | 100 | -1 | |
| 0.50:0.95 | large | 100 | 0.593 | |
| Average Recall | 0.50:0.95 | all | 1 | 0.6 |
| 0.50:0.95 | all | 10 | 0.683 | |
| 0.50:0.95 | all | 100 | 0.683 | |
| 0.50:0.95 | small | 100 | -1 | |
| 0.50:0.95 | medium | 100 | -1 | |
| 0.50:0.95 | large | 100 | 0.683 |
paper sheet : heatmap exploration and train to determine the color detection value of each pixel
Heatmap-based Object Detection and Tracking with a Fully Convolutional Neural Network
The visual digital turn: Using neural networks to study historical images
A Deep Residual Network with Transfer Learning for Pixel-level Road Crack Detection
A fully open-source framework for deep learning protein real-valued distances
A Decade Survey of Content Based Image Retrieval using Deep Learning